🧩 Dynamic Retriever for In-Context Knowledge Editing via Policy Optimization

This repository contains the official implementation of the EMNLP 2025 paper:

Dynamic Retriever for In-Context Knowledge Editing via Policy Optimization
Mahmud Wasif Nafee¹², Maiqi Jiang³, Haipeng Chen³, Yanfu Zhang³
¹ Rensselaer Polytechnic Institute (RPI), Troy, NY, USA
² Bangladesh University of Engineering and Technology (BUET), Dhaka, Bangladesh
³ College of William & Mary, Williamsburg, VA, USA

---

📄 Paper

Accepted at EMNLP 2025 (Main Conference, Poster)
📘 Read on arXiv

---

🧠 Overview

DR-IKE introduces a policy-optimized retriever that dynamically selects in-context examples for knowledge editing.
It achieves the best balance between edit success rate, paraphrase consistency, and retention rate across standard benchmarks such as CounterFact, ZSRE, and Wikidata-CF.

---

🚀 Getting Started

Installation

git clone https://github.com/wasifnafee/DR-IKE.git
cd DR-IKE
pip install -r requirements.txt

---

📚 Citation

If you find this repository useful, please cite our paper:

@misc{nafee2025dynamicretrieverincontextknowledge,
  title={Dynamic Retriever for In-Context Knowledge Editing via Policy Optimization},
  author={Mahmud Wasif Nafee and Maiqi Jiang and Haipeng Chen and Yanfu Zhang},
  year={2025},
  eprint={2510.21059},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2510.21059}
}